Vehicle Self-localization using 3D Building Map and Stereo Camera

Self-localization is one of the most important part in autonomous driving system. In urban canyon, the multipath and non-line-of-sight effects to GPS receiver decrease the precision of self-localization of the vehicle. More specifically, the lateral error is more serious because of the blockage of the satellites. However, the building on roadside could be the stable reference object for localization. Therefore, this paper proposes to use stereo camera and 3D building map to reduce the lateral error of positioning result. In our proposal, stereo camera is used to detect and reconstruct the building side view. Lateral distance between building and vehicle estimated by stereo camera is compared with 3D building map to rectify the lateral position of vehicle. In addition, this paper employs inertial sensor and GPS receiver to decide the longitudinal position of vehicle. The particle filter is used for the sensor fusion. The experiment is conducted in the center of Tokyo, Japan, which is a typical urban city scene with high density of tall buildings. It demonstrates that the proposed method could achieve sub-meter level accuracy in GPS difficult environments.